
Boxplot Variations in a Spatial Context:
An Omernik Ecoregion and Weather Example

By

Daniel B. Carr
Anthony R. Olsen
Suzanne M. Pierson
Jean-Yves P. Courbois



1. Introduction

This article presents three variations on boxplots for use in a spatial
context.  The three variations 
are linked micromap boxplots, linked micromap bivariate boxplots, and
angular boxplot glyphs.  
The specific examples show Omernik Level II Ecoregions.  The variables
chosen for illustration 
purposes are growing degree days and precipitation.

This article is closely related to papers by Carr et al (1998a, 1998b).  
The first paper provides a 
general description of linked micromap (LM) plots such as that in Figure
1b.  The second paper 
puts LM plots to work in describing Omernik Level II ecoregions.  It also
promotes LM plots as 
useful methodology in the KDD pattern discover process and as overviews
that provide an 
orienting context for drilling down to finer detail in extensive
hierarchically organized 
summaries.  This relatively brief article adapts graphics and text from the
second paper while 
focusing attention on boxplots.

In this paper Section 2 provides context for the graphics with a
description of Omernik 
ecoregions and the data sets.   Section 3 provides background on boxplots
and then presents both 
univariate and bivariate LM boxplots.  Section 4 motivates two angular
boxplot glyphs and 
presents the one that is more striking, the Portuguese man of war.  Section
5 closes with 
connections to other work and challenges for the future.

2.    Ecoregions and Datasets
       
While this paper focuses on graphics design, a scientific context exists
behind the examples: the 
presentation of summary statistics for ecoregions and the attempt to learn
about the intellectual 
structure that underlies ecoregion definition.  In this section we provide
background on 
ecoregions and associated datasets.   

2.1 Ecoregions

Ecoregions are a way of codifying the recognizable regions within which we
observe particular 
patterns or mosaics in ecosystems.  The general consensus is that such
ecological regions, or 
ecoregions exist.  However, disagreement in how the regions should be
constructed continues to 
be a scientific issue (Omernik, 1995).  In describing ecoregionalization in
Canada, Wiken (1986) 
stated:

Ecological land classification is a process of delineating and classifying
ecologically 
distinctive areas of the earths surface.  Each area can be viewed as a
discrete system 
which has resulted from the mesh and interplay of the geologic, landform,
soil, 
vegetative, climatic, wildlife, water and human factors which may be
present.  The 
dominance of any one or a number of these factors varies with the given
ecological 
land unit.  This holistic approach to land classification can be applied
incrementally 
on a scale-related basis from very site-specific ecosystems to very broad
ecosystems.

Within the United States two alternative approaches in the construction of
ecoregions are those 
developed by Omernik (1987, 1995) and Bailey (1995a, 1995b, 1998).  Each
constructs a 
hierarchy of ecoregions that corresponds to viewing the United States at
different scales.  
Omerniks approach is conceptually similar to that described by Wiken,
where the ecological 
regions gain their identity through spatial differences in the combinations
of the defining factors 
and which factors are important vary from one place to another and at all
scales.  Bailey (1998) 
develops ecological regions hierarchically.  First, he identifies
ecological regions of continental 
scale based on macroclimate, where macroclimates influence soil formation,
help shape surface 
topography, and affect the suitability of human habitation.  The continent
is subdivided with 
three levels of detail into domains, within domain divisions, and within
division provinces.  
Domains and divisions are based largely on broad ecological climatic zones
while provinces 
further sub-divide the divisions on the basis of macro features of the
vegetation.  Hence Bailey 
uses macroclimate as the controlling factor in the formation of ecoregions
while Omernik uses 
all available factors where the importance of the factors varies among
ecoregions.  Some 
scientists question whether enough is known to delineate ecoregions.  While
our knowledge is 
limited, others proceed on the basis that our approximations of the true
ecoregions will 
continue to improve as more information is gathered.

Our interest is not to define ecoregions or even to validate them.  We
simply believe that it is 
important to describe quantitatively some of the key characteristics
associated with ecoregions to 
gain a better understanding of how ecoregions have partitioned these
characteristics.   In view of 
the newsletter page size, we have elected to show Omerniks level II
ecoregions (Figure 1a) for 
the continental U.S. examples here.  This coarse level of aggregation
bothers some people, for 
example when they see the Mississippi Valley in the same ecoregion with the
Southeastern 
seacoast and Cape Cod.  Level III and level IV ecoregions are progressively
finer partitions of 
North America and are much more homogeneous in terms of familiar variables.
  

2.2 Data Sets

The data sets used in our graphics are readily available.  Substantial
thought and processing was 
involved in producing the data sets.  Thus this paper builds upon the work
of others.

Ecoregion delineation depends on climate as a major forcing function.  We
use nationally 
consistent climate data sets developed using PRISM (Parameter-elevation
Regressions on 
Independent Slopes Model).  PRISM, described by Daly et al (1994), is an
analytical model that 
uses point data and a digital elevation model (DEM) to generate gridded
estimates of monthly 
and annual climatic parameters.  PRISM models data from individual climate
stations to a 
regular grid through locally weighted precipitation/elevation regression
functions.  Orographic 
effects on precipitation and other climate parameters are well known and
are used to define the 
local regions.  PRISM has been shown to out perform other common spatial
interpolation 
procedures such as kriging.  The regular grid used is 2.5 minute by 2.5
minute latitude/longitude 
grid, i.e., a nominal 5 km by 5 km grid.  For our purposes having a regular
grid of climate data 
ensures that the entire area of an ecoregion is represented in any
statistical summary we 
construct.  PRISM data we use are based on 30-year normal climate data for
1961-1990.  Key 
annual parameters investigated are mean temperature, temperature range,
growing degree days, 
and total precipitation, and monthly precipitation. Further information on
PRISM and the data 
sets can be found at http://www.ocs.orst.edu/prism/. 

Figures 1 and 2 focus attention on just two PRISM variables, growing degree
days and 
precipitation.  The delineation of ecoregions depends not only on the
annual characteristics of 
climate but also on seasonal characteristics.  Thus monthly precipitation
is important to study.  
As indicated above many variables such as elevation, land cover, and soils
also relate to 
ecoregions.  A dated visual summary of level II ecoregion land cover
classes is available in Carr 
and Olsen 1996.  (A few  boundaries have changed.)  

Figure 3 in this paper concerns precipitation trends.  The figure makes
direct use of annual 
weather station precipitation data for the period 1961-1996.  The
precipitation data (72 Mbytes) 
and weather station location descriptors (10.2Mbytes) are available from
the National Climatic 
Data Center web site
http:www.ncdc.noaa.gov/ol/climate/onlye/coop-precip.html#Files.  

Our data processing classified weather stations into ecoregions.  We used
each stations most 
recent latitude and longitude (some change over time) to locate it in one
of 672 polygons that 
define the Level II ecoregions in the continental U.S.  We omitted stations
that had fewer than 24 
annual values over the 36 years.  Some 4075 weather stations satisfied the
criteria of being in an 
ecoregion and having sufficient annual values.  We then calculated trend
using Sens slope, 
which is the median of slopes computed using all pairs of available years
for each station.   This 
provides an estimate that is little influenced by an occasional extreme
value.   The angular 
boxplot glyphs in Figure 3 represent variation in 36-year trends at
different weather stations with 
the 21 Level II ecoregions.

3. LM Boxplots 

Figure 1b shows linked micromap univariate boxplots.  We briefly discuss
the basic LM plots 
elements: micromaps, linking labels, boxplots, and sorting.  For more LM
plot details see the 
previously mentioned papers.  

The micromaps on the left are caricatures of level II ecoregions.  The
caricatures simplify the 
polygon boundaries and reduce the 672 polygons for the continental U.S. to
44 polygons.  
Arguments can be more for going even further and, for example, removing
Cape Cod and the 
island in Lake Superior.

Omerniks Level II definition of ecoregions partitions the continental U.S.
into 21 ecoregions.  
The long labels for the ecoregions appear in Figure 1a.  Linking ecoregion
name to location is 
difficult in Figure 1a.  There are many ecoregions and some involve
disjoint areas.  The colors 
may be pleasant to view, but are not easy to discriminate.  For most
readers the colors do not 
have names, and this can make it much harder to remember the exact color
that is to be found on 
the map.  Plotting the ecoregion numbers on the map typically solves the
memory problem, but 
serial search is still involved.  With disjoint areas for the same
ecoregion, finding one number 
does not mean the search task has been completed.  The symbol congestion
and search problem 
gets worse with Level III ecoregions. 

Micromaps with color links provides a solution to the name and region
linking problem.  To find 
the polygons for Ecoregion 3 one simply looks for the red regions in the
top micromap.  The 
only difficulty concerns micromap caricatures with polygons so small that
they do not reveal 
their color.  The state map in Carr and Pierson 1996 enlarge small states
to make sure their color 
is visible.  In the current map caricature, Cape Cod (Ecoregion 9) is too
small to reveal its color.
The development of micromap caricatures involves compromise.

The use of integers as ecoregion labels is another compromise.  Descriptive
labels always have 
merit.  However, in this and other examples involving more statistical
summary columns, we 
choose to reserve space for the statistical graphics panels.  The integer
labels take up little space 
and become familiar with repeated usage. 

The color rectangle as a part of the label can be dropped in Figure 1b
because the boxplots 
provide the color.  For those experienced with LM plots, dropping the line
of colored rectangles 
simplifies appearance and complicates color linking only slightly.  We
retain the rectangles in 
the current context.  
 
The boxplot or box and whiskers plot is a well-known distributional
caricature that has now 
appeared in some grade school curriculum.  However, the choice of summary
statistics and the 
graphical representation are not universal (see McGill, Tukey and Larsen
1978; Frigge, Hoaglin, 
and Iglewicz 1989; Tukey 1993; and Carr 1994).  Figure 1b shows a five
number summary: the 
extrema, the 1st and 3rd quartiles and the median.  Another common choice
represents adjacent 
values and outliers rather than just extrema.  We note that the outlier
portion of this choice does 
not scale well for large samples from thick tailed distributions.   For
example if estimates 
followed a t-distribution with 3 degrees of freedom roughly 2.75 percent of
the estimates would 
be flagged as outliers on each end of the distribution.  Outlier
overplotting was problematic for 
precipitation in an alternative version of Figure 1b.  When outlier
overplotting is problematic a 
compromise caricature that we do not show uses adjacent values and adds an
outlier dot for a 
maximum or a minimum that is not an adjacent value.  The compromise uses a
filled dot to 
indicate multiple outliers thus hiding the details of outlier multiplicity
and location.

Our graphical representation follows the design of Carr (1994).  That is,
the thick rectangle 
extends from the 1st to the 3rd quartile, the thin rectangle extends to the
extrema, and the vertical 
median line extends outside the thick rectangle.  Comparison of adjacent
medians can typically 
be based on the judgement of horizontal distance between line endpoints. 
Using dots for the 
median (Becker and Cleveland 1993) is a viable choice since the area symbol
is easy to spot 
even though enclosed in a rectangle.  However, the comparison of
neighboring medians may not 
be as accurate since the direct distance between points is not the correct
distance to judge. 

The boxplot is suitable for describing a batch of data.  The batch of data
here reflects variation 
due to changing spatial location.  However, one would like the items in a
batch to be as 
comparable as possible.  This is not the case here because the cells are
defined using equal angle 
grids and not equal area grids.  Carr et al (1997) question the use of
equal angle grids in 
environmental sciences and in NASAs level 3 satellite products.  They
promote the use of
discrete global equal area grids.  While transformations can be made so the
North Pole is not as 
wide as the equator, changing to equal area grids for reasons of
comparability or polar modeling 
adds complexity and uncertainty.  We choose to avoid the complexity of
re-gridding the data for 
this article on boxplots.  A consequence is that in large north to south
ecoregions, such as 
Ecoregion 9, the northern portion is over represented in the boxplot due to
smaller area grid cells.  
If the North is cooler than the South the growing degree day percentiles
will be shifted a bit 
toward cooler values.  

The notion of growing degree days many not be familiar, but the calculation
is straight forward.  
If the average daily temperature is over 50 degrees Fahrenheit, the day
counts.  The degrees 
accumulated for each day that counts is the average daily temperature minus
50 degrees.   As a 
calibration case suppose a grid cell in Florida had a daily average value
of 75 degrees, each day 
for the whole year.   The growing degrees days would then be 365 * (75-50)
= 9125.  To provide 
easier to remember two digit numbers, the scale for Figure 1b shows growing
degree days 
divided by 100.   
  
The precipitation data for Figure 1b has a thick right tail.  A logarithmic
transformation helps to 
pull in the tail, and provides better resolution for small precipitation
values.  The choice here was 
to use log base 2, since powers of 2 are familiar to many scientists (see
Cleveland 1985).     

Sorting helps to bring out patterns in LM plots.  Carr and Olsen (1996)
discuss sorting methods 
and show ecoregion examples.  In Figure 1b sorting by increasing growing
degree days arranges 
the ecoregions in the micromap panel sequence so that northern ecoregions
tend to appear in the 
top panels and southern ecoregions appear in the bottom panels.  This
matches common 
expectation.  Some of the north to south anomalies, such as Ecoregion 20,
are easier to 
understand when elevation boxplots are included.   While pattern details
may be well-known to 
experts, those new to the field may delight in seeing patterns that confirm
their limited 
knowledge and in finding anomalies.  As an interpretation reminder, the
patterns in growing 
degree days refers to the thirty year average and not to a particular day
or year.

Sorting Figure 1b by increasing precipitation reveals a strong West to East
pattern.   With the 
exception of Ecoregion 4 that contains the Pacific Northwest rain forest,
the high precipitation is 
in the East.   A plot is available on our web site.    

The juxtaposed plots make it tempting to look at the growing degree day and
precipitation 
medians to see if there is a relationship.  The first impression is that if
there is a relationship, it is 
weak.  Of course juxtaposed univariate plots, such as bar plots, dot plots
or boxplots, provide a 
poor way to look for a functional relationship even if one variable is
sorted. While perhaps 
unknown to popular press, the scatterplot provides the standard for
assessing a functional relation 
between two variables.  We omit this scatterplot and proceed to represent
the bivariate data for 
all grid cells and not just the univariate medians.  

The current application with 481,475 grid cells warrants the use of density
estimation as part of 
the graphic representation process.  Scott (1992) provides methods for both
density estimation 
and graphical representions that emerged from years of research with
binning and other density 
estimation methods.  Proponents of binning for dealing with large data sets
also include Carr et 
al (1987).  Their space-time example included point in polygon selection on
a map to defined 
two subsets, representation of density differences in a scatterplot matrix,
and a representation of 
a temporal mismatch.  Bivariate binning methods scale well in the number of
cases.  Carr (1998) 
shows a scatterplot matrix representing over  billion point pairs.  Such
density estimates can 
vary over many orders of magnitude.  Appreciating the density surface
details is not necessarily a 
trivial task.  For an overview we seek something far simpler and turn to a
bivariate boxplot.

Chances are that many people have proposed bivariate boxplots.  So far
bivariate boxplots have 
not seemed to catch on, likely due to lack of promotion.  Possibly
disagreement on details is 
behind this lack of promotion.  Certainly there are issues concerning what
generalization is most 
appropriate for the 1st, 2nd, and 3rd quartiles.  (See Small [1989]
concerning bivariate median 
options that were available years ago.)  Perhaps it is better to show
density modes and two alpha 
contours. Our attitude is to get on with the graphics.  We illustrate the
approach of Carr (1991) 
but are quite willing to substitute other methods.   

Carr (1991) binned the data using hexagon grids to speed the processing. 
He directly used the 
high density cells containing 50 percent of the data in place of the
interquartile range.  To obtain 
a median he smoothed the high density cells (an optional step), and then
eroded the cells using 
gray level erosion. The gray level erosion process removes counts from the
cells proportional to 
exposed surface area.  The last cell eroded provides location of the median
in the binned density 
representation.  (S-PLUSTM now provides the needed tools for hexagon
grids.)  This simple 
procedure generalizes to other regular grids in two and higher dimensions. 
With translucence or 
the see through contour representations in Scott (1992), showing 3-D
boxplots is not a problem.  

If one were to use the 50% high density region in univariate boxplots
rather than the interquartile 
range, comparison becomes harder because the high density region can be
disjoint.   Carr (1991) 
addresses the increased difficulty of comparison in the bivariate context
by developing a 
difference plot.  Cells in two gray levels distinguish the two mismatched
portions of the high 
density region.  An arrow shows the change in the median.  The example
interleaves difference 
plots between bivariate boxplots in an age and sex conditioned two-way
layout of bivariate 
boxplots.  With more space we could interleave the bivariate difference
plots in the linear 
sequence shown in Figure 2.  

Figure 2 shows a bivariate boxplot of the combined data in the lower right
corner.  The purple 
shows the support of the bivariate pairs, (30-year average precipitation on
a log scale, 30-year 
average growing degree days).  The yellow cells are the 50% high density
cells and the black cell 
is the median cell.  The positive trend of the high density cells contrasts
strongly to the negative 
slope suggested by the region of support.   The latter is all one would see
in an overplotted view.  
How useful it is to study this aggregate plot can be argued.  Scientists
tend to be interested in 
detail.  When scientists see a plot splitting into non-homogeneous pieces
as in Figure 2 they 
naturally want to study the pieces.  Nonetheless overviews are often
valuable for providing a 
frame of reference as one approaches the pieces.      

We mention several patterns in Figure 2, but the existence of Level IV
ecoregions suggests that 
scientists would prefer to focus attention on more highly stratified
ecoregion views.  Already 
suggested was the fact the variation of individual ecoregions is generally
much smaller than that 
of the composite.  The small amount of yellow in the Ecoregion plots
indicates a high 
concentration of bivariate values within tight domains.  Several boxplots
suggest that they are 
mixtures of tight patterns.   Few bear much resemblance to a bivariate
normal distribution.  A 
moderately elliptical exception, Ecoregion 10, shows no yellow so the cell
with yellow 
overplotting black median dot contains over 50 percent of the observations.
 Panels 15, 18, and 
16 (top to bottom order) catch the eye with relatively big bivariate
support regions (gray) that 
have negative slopes.  The high density yellow cells reflect the negative
slopes as well.  The high 
density cells in Region 16 show a bifurcation that motivates further
investigation and suggests 
possible subdivision of the ecoregion.  (Carr et al 1998b note that Region
16 is homogeneous in 
terms of land cover). Regions 10 and 9 show high positive slopes.  This
suggests that most of the 
variation is in terms of growing degree days and that growing degree days
is not major factor in 
their definition.   Note that both regions cover a large range of latitude.
 In general the growing 
degree day variation in the bivariate boxplot appears associated with
latitude variation in the 
micromaps.  This motivates putting growing degree days on the y axis.

Figure 2 uses the univariate growing degree days as the basis for sorting,
but the micromap 
layout is different than in Figure 1.  Due to the size of bivariate
boxplots, Figure 2 shows four or 
fewer ecoregions per micromap.  The vertical list of names and color links
is absent since they 
are mostly redundant with the names and color tags in each bivariate
boxplot panel.  The list 
would clarify the bivariate boxplot order that is left to right, top to
bottom.  Careful comparison 
of bivariate median values for growing degree day values against the
reference grid indicates the 
bivariate and univariate ranking are in reasonable agreement.  Other
sorting can bring out other 
patterns.  

4.  Angular Glyph Boxplots and a Precipitation Trend Example
Carr and Pierson (1996) favored LM plots over choropleth maps in terms of
representing 
statistical summaries more faithfully.  Their reasons include better
perceptual accuracy of 
extraction and easy visual representation of uncertainty.  However, classed
choropleth maps are 
not the only way to represent spatially indexed statistical summaries. 
Glyph plots provide a 
viable alternative for representing summaries so they deserve
consideration.  
Carr et al (1998b) propose two angular glyph boxplots but illustrate only
one.  Figure 3 shows 
the more controversial glyph that we call the Portuguese man of war.   The
glyph extends the 
trend glyph design of Carr, Olsen, and White (1992).   In their trend glyph
design, ray angle 
represents the sulfate deposition trend and the straight edges of a circle
sector represent 90 
percent confidence bounds.  A local scale, a circle with lines at regular
angles, provides the basis 
for judging ray angles.  We continually attempt to put Cleveland and
McGills (1984) guidance 
concerning perceptual accuracy of extraction into practice.  They promote
representations that 
use position along a scale as being the best.   Our encoding uses position
along a scale, albeit an 
angular scale.  

Then boxplot glyph shown in Carr et al (1998b) uses the circle sector to
represent the interval 
from first to third quartile.  Two shorter length rays (but longer than the
reference circle) encode 
the extrema.  The Portuguese man of war glyph shortens the extrema rays to
the length of the 
reference circle and cuts away the circle sector between extrema rays.  As
can be seen in Figure 
3, the glyph also removes the arc between lines showing the first and third
quartiles.  The gestalt 
is striking.  Some symbols are aligned and some twist.  The departures from
symmetry about the 
horizontal zero trend line and about the median line are clear.  The
gestalt is so striking that it 
may impede people looking at the detail.  Also cutting away part of the
reference circle makes it 
a bit harder to judge angles.  Thus this Portuguese man of war glyph may
lose in careful testing 
for perceptual accuracy of extraction.   

The angular limits for the boxplot glyph in Figure 3 are from  135 degrees
to 135 degrees.   
There is some advantage in restricting the limits to 90 degrees so, for
example, the largest local y 
coordinate indicates the largest increase.  The extension here is to
increase angular resolution.  
The legend scale has limits of  -.91 and .91 inches per year.  This implies
a change of roughly 32 
inches over the 36 years.  The scale is defined symmetrically so both
extremes are not 
necessarily achieved.  However, a study of the glyphs indicate that this is
nearly so.  Time series 
plots for individual stations confirm such substantial change.   

The existence of multiple polygons for the same ecoregion complicates
examination of Figure 3.  
The same glyph appears for each polygon that comprises a given ecoregion. 
This gives more 
visual weight to ecoregions with multiple polygons.  If unequal visual
weight is given, a more 
reasonable choice is to favor large area ecoregions.  The glyph design
itself could be modified to 
indicate the number of weather stations with each ecoregion.
 
Glyph plots have some problems.  We have slightly revised the GIS-provided
representative 
point for each polygon to eliminate glyph overplotting in Figure 3.  Symbol
congestion control 
becomes increasingly problematic as the number of polygons increase, for
example with level III 
ecoregions.  (The trend glyph plot in Carr, Olsen and White [1992] was
restricted to a regular 
hexagon grid to control symbol congestion.)  The glyphs provide a spatial
overview, but in this 
gray level example do not connect strongly to ecoregions and their names. 
LM plots provide a 
strong summary to name connection and avoid multiple glyphs, placement and
symbol 
congestion problems.  Also the straight line scale boxplot is a bit easier
to assess than an angular 
scale boxplot.  Still, the gestalt spatial impression of glyph plots has
value. 

5.   Connections and Future Challenges 

The article builds upon previous work by many people.  Carr et al (1998b)
cite many that 
inspired our development of LM plots.   Our shortened list here includes
Cleveland and McGill 
(1984), Cleveland (1985), Monmonier (1988, 1993), Tufte (1990, 1993) and
Kosslyn (1994).  
Our scholarship in regard to bivariate boxplots is at best dated.  We
welcome references to the 
work of others and constructive comments about better designs.  

The Splus script files we used to produce the graphics are available by
anonymous ftp to 
galaxy.gmu.edu.  The subdirectory is pub/dcarr/newsletter/boxplots.  The
contexts include the 
statistical summaries used in the graphics but not the prior data.  For
more information on the 
data or on algorithms such as gray level erosion in 3-D and 4-D please
contact us directly.  There 
is also a web site,
www.galaxy.gmu.edu/~dcarr/graphgall/ecoregions/ecoregions.html that 
contains a growing number of Omernik level II and level III ecoregion
examples.  

We want to make a point relative to our first presentation (Olsen et al
1996) of Omernik 
ecoregions.  The 4 x 8 foot poster example provided a beginning description
of  Omerniks Level 
III ecoregions.  This poster included over one hundred micromaps, boxplots
of ecoregion digital 
elevations and even Lovelands 8 million pixel map of 159 land cover
classes for the continental 
U.S.  (See Loveland et al 1995).  While large high-resolution graphics that
integrate different 
views would seem to provide a powerful tool in the attempt to understanding
large data sets, 
such graphics fall outside the mainstream.   Todays emphasis is on web
graphics.  Tufte  
(http://www.clbooks.com/nbb/tufte.html) comments that, the problem with
the web is that it is 
low resolution in both space and low in time.  As one way to appreciate
this statement, note that 
a typical workstation monitor shows only 1024 * 1280 pixels (roughly 1.3
million pixels).   Thus 
without reducing resolution, Lovelands map requires over six complete
screens.  Something is 
lost in comparison to high quality printed maps.  There is no denying the
power of human 
computer interface methodology such as logical pan and zoom, but the
challenge of large data 
sets and complex conceptual structures motivates use of all powerful tools
including good old 
human pan and zoom of large high quality graphics.

The task of developing and presenting statistical overviews for massive
data sets is a big 
challenge. The environmental science community has worked on the task
longer than many 
disciplines.  How does one represent environmental facts and knowledge? 
Two recent book 
length overviews, Stone et al (1997) and Wahlstrom, Hllanaro, and Maninen
(1996), involved 
the work of many people, including in one case layout specialists and
graphics designers.  The 
integration of photographs, text, and statistical graphics serves as an
inspiration.   Our challenge 
is to learn when others have blazed the trail, to make the improvements
when we see the 
opportunity and to adapt the methods to other applications.   Perhaps an
even harder challenge is 
to see the patterns in important data not collected and take action.     

6.  Acknowledgements

S-PLUS is a registered trademark.

EPA funded the majority of the work behind this paper under cooperative
agreements No. 
CR8280820-01-0 and No. CR825564-01-0.  Additional federal agencies, BLS and
NCHS, 
supported some facets of this work.  The article has not been subject to
review by BLS or NCHS 
so does not necessarily reflect the view of the agencies, and no official
endorsement should be 
inferred.  It has been subjected to EPAs peer and administrative review
and approved for 
publication.  The conclusions and opinions are solely those of the authors
and are not necessarily 
the views of the Agencies.  

7.  References

Bailey, R.G.  1995a.  Description of the ecoregions of the United States.
Misc. Publ. No. 1391 
(rev.), USDA Forest Service, Washington, DC.  

Bailey, R.G. 1995b.  Ecosystem Geography. Springer-Verlag, New York.
Bailey, R.G. 1998. Ecoregions map of North America: Explanatory Note. Misc.
Publ. 1548, 
USDA Forest Service, Washington, DC.  
Becker , R. A. and W. S. Cleveland.  1993.  Discussion of Grapical
Comparison of Several 
Linked Aspects:  Alternative and Suggested Principles, Journal of
Computational and 
Graphical Statistics, Vol. 2. No. 1, pp 41-48.

Carr, D. B.  1991.  "Looking at Large Data Sets Using Binned Data Plots." 
Computing and 
Graphics in Statistics, eds. A. Buja and P. Tukey, pp. 7-39. 
Springer-Verlag, New York, New 
York.  

Carr, D. B.  1994.  "A Colorful Variation on Boxplots" Statistical
Computing & Graphics 
Newsletter, Vol. 5, No. 3, pp. 19-23.

Carr, D. B.  1998.   Multivariate Graphics, Encyclopedia of
Biostatistics,  Eds. P. Armitage and 
T. Colton, Vol. 4, pp. 2864-2886.

Carr, D. B. and A. R. Olsen 1996.  Simplifying Visual Appearance By
Sorting:  An Example 
Using 159 AVHRR Classes,  Statistical and Statistical Graphics Newsletter,
Vol. 7 No. 1  pp 
10-16.
Carr, D. B., A. R. Olsen, and D. White.  1992.  "Hexagon Mosaic Maps for
Display of 
Univariate and Bivariate Geographical Data."  Cartography and Geographic
Information 
Systems, Vol. 19, No. 4, pp. 228-236,271.

Carr, D. B., A. R. Olsen, J. P. Courbois, S. M. Pierson, and D. A. Carr. 
1998a.  Linked 
Micromap Plots: Named and Described, Statistical Computing & Graphics
Newsletter, Vol. 
9 No. 1 pp. 24-32.

Carr, D. B., A. R. Olsen, S. M. Pierson, and J. P. Courbois. 1998b.  Using
Linked Micromap 
Plots To Characterize Omernik Ecoregions, Special Issue on Statistical
Mining And Data 
Visualization in the Atmospheric Sciences, T. Brown and P. Mielke Eds.  
Data Mining and 
Knowledge Discovery.  Submitted.

Carr, D. B., R. J. Littlefield, W. L. Nicholson, and J. S. Littlefield. 
1987.  Scatterplot Matrix 
Techniques For Large N.  Journal of the American Statistical Association
82 (398):424-436.
Carr, D. B., R. Kahn, K. Sahr, and A. R. Olsen.  1998a.   ISEA Discrete
Global Grids, 
Statistical Computing & Graphics Newsletter, Vol. 8 No. 2/3 pp. 31-39.

Carr, D. B. and S. Pierson.  1996.  Emphasizing Statistical Summaries and
Showing Spatial 
Context with Micromaps,   Statistical Computing & Graphics Newsletter,
Vol. 7, No. 3, pp. 16-
23.

Cleveland, W. S.  1985. The Elements of Graphing Data, Monterey, CA: 
Wadsworth;

Cleveland, W. S  and R. McGill.  1984.  Graphical Perception:  Theory,
Experimentation, and 
Application to the Development of Graphical Methods.  Journal of he
American Statistical 
Association, 79:531-554.

Daly, C., R.P. Neilson, and D.L. Phillips (1994) A statistical-topographic
model for mapping 
climatological precipitation over mountainous terrain. Journal of Applied
Meteorology, 33, 140-
158.

Frigge, M., D. C. Hoaglin, and  B. Iglewicz,  1989.  Some Implementations
of the Box Plot,  
The American Statistician, 43, pp. 50-54.
Kosslyn, S. M.  1994.  Elements of Graph Design, New York, NY: W. H.
Freeman and 
Company.
Loveland, T. R., J. W. Merchant, B. C. Reed, J. F. Brown, D. O. Ohlen, P.
Olson, J. Hutchinson.  
1995.  Seasonal land Cover Regions of the United States.  Annals of the
Association of American 
Geographers.   Vol. 85, pp. 339-355.

McGill, R.,  J. W. Tukey, and W. A. Larsen.  1978.  Variation of
Boxplots,  The American 
Statistician, 32, 12-16.  

Monmonier, M. 1988.  Geographical Representations in Statistical Graphics:
 A Conceptual 
Framework  American Statistical Association 1988 Proceedings of the
Section on Statistical 
Graphics.  American Statistical Association, Alexandria VA.  pp. 1-10.

Monmonier, M.  1993.  Mapping It Out.  The University of Chicago Press,
Chicago, IL.
Olsen, A. R., D. B. Carr, J. P. Courbois, S. M. Pierson.  1996. 
Presentation of Data in Linked 
Attribute and Geographic Space.  Poster presentation,  ASA Annual Meeting,
Chicago, Il.

Omernik, J.M.  1987. Ecoregions of the conterminous United States. Annals
of Association of 
American Geographers, 77(1), 118-25.

Omernik, J.M.  1995. Ecoregions: a spatial framework for environmental
management. In 
Biological Assessment and Criteria: Tools for Water Resource Planning and
Decision Making, 
W.S. Davis and T.P. Simon (eds.), Lewis Publishers, Boca Raton, Florida. 
pp. 49-62.

Omernik, J.M. and Bailey, R.G.  1997.  Distinguishing between watersheds
and ecoregions. 
Journal of The American Water Resources Association, 33(5), 935-49.

Scott, D. W.  1992.  Multivariate Density Estimation, Theory, Practice and
Visualization.  John 
Wiley and Sons, Inc.  New York.  

Small , C. G. 1989.  A Survey of Multidimensional Medians.  Statistical
Technical Report Series 
Stat-89-08, Department of Statistics and Actuarial Science, University of
Waterloo, Waterloo, 
Canada. 

Tufte, E. R. 1983.  The Visual Display of Quantitative Information,
Cheshire, CT: Graphics 
Press.
Tufte, E. R. 1990.  Envisioning Information, Cheshire, CT: Graphics Press.
Tukey, J. W.  1993.  Graphic Comparisons of Several Linked Aspects:
Alternatives and 
Suggested Principles. , Journal of Computational and Graphical Statistics,
vol. 2. No. 1, pp 1-
33.
